Toward Human-Like Sentence Interpretation--a Syntactic Parser Implemented as a Restricted Quasi Bayesian Network--

摘要

Most sentences expressed in a natural language is ambiguous. However, human beings effortlessly understand the intended message of the sentence even when a computer program finds out countless possible interpretations. If we want to create a computer program that understands a natural language in the same way as human beings do, a promising way would be implementing a human-like mechanism of sentence processing instead of implementing a "list exhaustively then select" method. By the way, it is highly likely that human's language ability is realized mostly by the cerebral cortex, and recent neuroscientific studies hypothesize that the cerebral cortex works as a Bayesian network. Then it should be possible to reproduce human's language ability using a Bayesian network. Based on this idea, we implemented a syntactic parser using a restricted quasi Bayesian network, which is a prototyping tool for creating models of cerebral cortical areas. The parser analyzes a sequence of syntactic categories based on a subset of combinatory categorial grammar. We confirmed that the parser correctly parsed grammatical sequences and rejected ungrammatical sequences.